WHAT IS POOR MAN’S EYE TRACKING GOOD FOR?

Carsten Ullrich Dieter Wallach Erica Melis

DFKI GmbH Digital Media Department DFKI GmbH
Stuhlsatzenhausweg 3 University of Applied Stuhlsatzenhausweg 3
D-66123 Saarbrücken, Sciences D-66123 Saarbrücken,
Germany Kaiserslautern, Germany Germany
+49 681 302 5370 [email protected] +49 681 302 4629
[email protected] [email protected]

ABSTRACT In this paper, we describe a new methodology extra effort [3] or they rely on the user’s compliance to
for tracking a user’s areas of interest on a computer screen. specific experimental instructions [6].
In contrast to cost- and analysis-intense eye-tracking studies, 2. DFK EYE : A POOR M AN ’S E YE T RACKER
our approach operates on a software-only basis that requires
Hardware eye trackers use infrared light reflections on the
no additional hardware. In an exploratory study, we
user’s cornea to track the fixations of participants. In this
compared our software-based tracker with hardware eye
paper, we argue that trailing a user’s area of interest can be
trackers with respect to the types of recorded data.
closely approximated by tracking the user’s mouse
Interestingly, our software-based «poor man's» eye tracker
movement in a proper setup [1], [2].
allows the collection of data that would pose several
problems for hardware eye tracker. In this paper, we describe When using our DFKeye, the screen is divided into several
a study in which we compared whether both approaches regions. Text areas, Figures, information containers or
yield similar data. A first evaluation of our results clearly arbitrary other user interface regions located within these
indicates high correlations between the software eye tracker regions are hidden from the user by blurring their content.
and the hardware system. To make the respective region’s content visible, the user
focuses it by moving the mouse pointer into this area. As
Keywords Byrne et al. [1] discuss, there is evidence that users tend to
Usability Testing, Eye Tracking, Empirical Methods, move the mouse pointers to the regions where they are
Toolkits gazing at. Whenever a user enters or leave a certain region,
DFKEye records the respective start and end times. From a
user’s point of view, the visibility of a region is removed
after a parameterizable time slot and its content is hidden
1. I NTRODUCTION again. To regain focus, the user needs to move the mouse
In his article “What is Eye Tracking Good for?” Schroeder again.
[5] enumerated the “Can” and “Can’t” of hardware eye 3 . W HAT I S P OOR M A N ’ S E YE T RACKING
trackers. At the German Research Center of Artificial G OOD FOR?
Intelligence in Saarbrücken we have developed a software-
There are a several reliably functioning hardware eye-trackers
based (poor man’s) eye tracker that overcomes most of the
on the market — so why can there be a need for a poor
“Can’t”s of hardware eye trackers that Schroeder discussed,
man’s eye tracker such as DFKeye? Besides the
while still offering almost the same advantages. DFKeye
disadvantages of the enormous costs for buying and
was initially developed for the use in the eLearning
maintaining a hardware eye tracker, we also found
environment ACTIVEMATH [4], however we recently
restrictions when applying it in research. On a general level,
started to explore its scope and applicability as a substitute
Schroeder [5] enumerates the types of data that can be
for hardware eye trackers with arbitrary interfaces.
gathered using hardware eye trackers. What about a software-
based system like DFKEye? According to Schroeder,
hardware-based eye-trackers:
® Can tell whether users are even looking at the
screen. Since DFKEye is not actually tracking
gazes, but focuses on the user’s behavior, it cannot
tell whether a user is just passively sitting in front
of the screen, or focusing any areas on it. However,
as soon as a user wants to use information on the
Although we know of two similar approaches to DFKeye, screen, her mouse movements will indicate her
their application presuppose a somewhat limited and areas of interest. In this sense, although DFKEye
artificial experimental setup since they either need cannot tell whether the user is not looking at the
participants to perform significant and cognitive demanding
 screen, we can certainly tell when she is and which procedure often takes almost as long as the
information she is requesting. experiment the eye tracker is actually used for.
® Tell whether users are reading or scanning. Even Depending on the type of hardware eye tracker
though we cannot distinguish between reading and (head-mounted vs. remote), participants need to be
scanning on the grain size of words, we can gather instructed not to move their heads too much. This
evidence for reading vs. scanning at the region requirement implies a rather uncomfortable
level by inspecting the focus times of the different situation for the participants and restricts the
regions. Short, frequently varying foci might duration of an experiment that makes use of eye
indicate whether a user is scanning for some item, tracking. In contrast, DFKEye does neither need a
whereas long, stable foci suggest that the user is calibration procedure, nor specific instruction
reading. Since the regions on the screen are only regarding head movement, and can thus even be
blurred, but not blackened, the user is still able to applied remotely via the Web. With only a few
see the overall layout of screen and has thus exceptions, most eye tracking studies tested only a
enough visual cues for finding relevant areas on the very small number of participants, or relied on
screen. single case studies — due to its effortless
applicability DFKEye promises to conduct studies
® Compare user scan patterns. This analysis requires with a broad number of participants.
the comparison of single user data, which both the
hardware- and the software-based eye trackers yield. DFKeye can be easily adapted to a broad range of
experimental requirements: Three different variants of hiding
® Learn the relative intensity of a user’s attention to the elements within a region have been implemented:
various parts of an interface. Information about
the distribution of a user’s attention to different ® cover (covers the respective completely),
screen areas, directly arises from the division of the ® fade (changes the font colour to approximately
screen into distinct regions. For each region match the background colour),
DFKEye registers when and how long the user ® and zoom (changes the font size to miniscule).
looked at it.
DFKEye allows the adjustment of the delay between
The comparison above indicates that our behavior-based entering a region with the mouse pointer and unhiding it, as
approach provides a functionality that allows to address well as setting the time after which the region is hidden
research questions that previously required expensive again when the mouse pointer is not moved.
hardware eye-trackers. It is not our intention to propose
DFKEye as a substitute for hardware-based system, DFKeye is realized as a Javascript application that uses
however, we found several questions that are not answered onMouseOver/on-MouseOut event handlers on HTML DIV
by the use of hardware eye trackers. DFKEye can, for elements to catch the mouse movements. Time information
example: is send to a servlet (java server application) immediately
after a region is blurred. All data is stored in an XML
® Let you know whether users actually see something format and can thus be easily transformed to the required
on the screen. A hardware eye tracker cannot input format easily using XSL-stylesheets. DFKEye is
distinguish between users studying the region they implemented in strict conformance to the W3C standards,
are looking at and users who simply stare at them, and is usable in all W3C-compliant browsers (Mozilla,
not perceiving the respective content. However, Netscape6) and in IE5. Currently, we are preparing a
since DFKEye blurrs the content of a region after a DFKEye-kit that can be easily integrated into existing web-
certain period of time, a. participant who wants to based systems.
continue studying the region, needs to move the
mouse in order to make the region’s content visible 3 . 1 Comparing Both Methodologies
again. As argued in the last paragraph, there are obviously several
® Prove that users didn’t see something. Peripheral advantages associated with the use of DFKEye in
vision cannot be tracked by a hardware eye tracker. comparison to a hardware-based tracker. However, the
In a setting that uses DFKEye, however, regions central question is whether DFKEye and a real eye tracker in
that are not focused cannot be perceived by a user fact yield similar data patterns when applied in a study. To
since they are blurred and their content therefore evaluate this question, we conducted a study targeted to
not accessible. analyze the relative strength of visual cues, using a hardware
eye tracker and DFKEye. Although data analysis is still
® Test everybody. Hardware eye trackers are difficult underway, our initial results are encouraging: To allow for a
to calibrate, to keep calibrated, and are sometimes direct comparison of the data gathered, we contrasted the
not usable with some participants at all, whereas data on the level of screen regions which were defined for
DFKEye requires users only being able to move a the hardware-based tracker (SMI) and for DFKEye. The
mouse. In our opinion, this amounts to be the correlation coefficient of the data gathered with the two
major advantage of DFKEye: Even when testing systems fall within the range between .73 and .94 and thus
completely cooperative participants, the calibration
indicate that the hardware-based eye tracker and DFKEye [2] Chen, M. C., Anderson, J. R., and Sohn, M. H.: What
obviously yielded similar results in this study. can a mouse cursor tell us more? Correlation of
In our opinion, DFKEye offers an easy-to-use, inexpensive eye/mouse movements on web browsing. Proceedings of
and reliable alternative to hardware eye tracking. Since Computer Human Interaction (CHI) 2001, p. 280-281,
DFKEye can easily be integrated into existing web-based 2001.
applications such as eLearning frameworks [4], it allows the [3] Egner, Itti and Scheier. Comparing attention models
use of evidence for a user’s regions of interest in domains with different types of behavior data, Investigative
where hardware-based eye trackers would not be applicable. Ophthalmology and Visual Science (Proc. ARVO 2000),
Vol. 41, No. 4, p. S39
4. A CKNOWLEDGEMENTS
We thank Michael Dietrich for his work for the [4] Melis, E., Büdenbender, J.,Andres, E.,Frischauf, A.,
implementation of various versions of DFKEye and Goguadze, G., Libbrecht, P., Pollet, M., and Ullrich, C.
ACTIVEMATH: A generic and adaptive web-based
Thorsten Moritz for conducting the evaluation study. learning environment. Artificial Intelligence and
5. R EFERENCES Education, 12(4), 2001.
[1] Byrne, Michael D., Anderson, John R., Douglass, Scott [5] Schroeder, Will: What is eye-tracking good for?
& Matessa, Michael: Eye Tracking the Visual Search of http://www.uie.com/eyetrack2.htm
Click-Down Menus. Proceedings of the SIGCHI [6] Wilhelm, T., Yom, W., and Berger, S. Site-Covering
conference on Human factors in computing systems: the eine innovative Methode zur Erfassung der
CHI is the limit, p. 402-409, 1999. Informationsaufnahme und des Entscheidungsverhaltens
auf Webseiten, Planung & Analyse, April 2002